Hi, I'm Achref Jarray
Senior AI Engineer and computer vision engineer building football performance systems, edge AI workflows, and agentic RAG solutions across sports, defense, and research.
Achref Jarray
Senior AI Engineer · Computer Vision Engineer
About Me
Senior AI Engineer with 4+ years of experience delivering production-grade AI systems across defense, sports analytics, and healthcare domains.
My work focuses on designing and deploying intelligent systems that combine computer vision, large language models, and retrieval-augmented generation to extract actionable insights from video and document data. I specialize in building end-to-end AI pipelines for real-time object detection, multi-object tracking, and natural-language interaction, enabling domain experts to query systems conversationally and receive context-aware responses.
My experience includes developing offline-first AI architectures for edge deployment, implementing feedback-driven model improvement loops, and constructing knowledge graphs for semantic reasoning. I earned a top-3 finish in an international satellite imagery competition and co-authored a peer-reviewed publication on deep learning for small-object detection under challenging conditions. Throughout, I prioritize systems that are accurate, interpretable, and aligned with end-user operational workflows.
Education
Research master's in Information System Techniques, ENIT
Nationality
Tunisian
Languages
English, French
Experience
Senior AI Engineer
- Led Football AI System (D-Fine, Norfair/DeepSORT, action recognition), 75% faster (10min → 2.5min per 30s video)
- Achieved 80–95% tracking accuracy with zero identity swaps via ReID-enhanced modes during high-intensity matches
- Built MCP AI Assistant (PostgreSQL/Neo4j), sub-3s queries, 500+/day, 94% satisfaction
- Agentic RAG + GraphRAG on Neo4j, 88% autonomous medical Q&A, 70% less reporting time
- MLflow CI/CD, 3-week → 3-day cycles, +18% F1, +22% precision
- QA lead: <2% false-positives, 98% on-time, 65% fewer defects
Master's Thesis, AI Research Internship
- Enhanced YOLO World with DCNv3, Coordinate Attention & AMMF modules + SAHI, +23% mAP (68.4% → 84.1%), +31% small-object recall
- 57% fewer false positives in fog/night conditions on FLAME dataset
Lieutenant, AI Engineer
- Built Chrome Extension (Manifest V3) for classified document Q&A & summarization, 65% faster review, 94% accuracy, 500+ reports/month
- Real-time edge video analysis: 30 FPS, <45ms latency, 91% anomaly detection, zero false negatives in live drills
- Offline-first architecture, 99.8% uptime over 30-day zero-connectivity deployments, <2% data drift
- INT8 quantization + ONNX Runtime operator fusion on constrained edge hardware
AI Engineer Internship
- Improved YOLOv5 with custom augmentation pipeline + Optuna optimization, 90% precision, 87% recall (F1: 88.5%), 42% fewer false alarms
- Real-time inference: 45 FPS on edge GPUs, <30ms latency, 24/7 autonomous airspace monitoring
Skills & Technologies
Programming & Tools
AI & Computer Vision
Data, Graphs & MLOps
Delivery & Prototyping
Education
Research master's in Information System Techniques
National Engineering School of Tunis (ENIT) 2023 - 2024Geomatics Engineer
Borj l'Amri Aviation School (EABA) 2018 - 2021Preparatory Cycle: Mathematics and Physics
Borj l'Amri Aviation School (EABA) 2016 - 2018Projects & Research
Aerosol Optical Depth Estimation
Achieved 3rd place in Solafune's AOD Estimation competition with 0.9896 public and 0.9893 private scores. Fine-tuned a Vision Transformer on Sentinel-2 data.
Football Performance AI System
A production football analytics platform combining object detection, multi-object tracking, action recognition, and an MCP-based AI layer connected to PostgreSQL and Neo4j.
Automated Drone Inventory System
Real-time aerial inventory analysis using autonomous drones with stream processing, deep learning-based OCR, and object detection. Achieved 95% detection accuracy.
Coaching Training Session Planner
AI-driven training session planner for football coaches. Built with web scraping, User Graph RAG, FastAPI, Ollama, and a reranking mechanism.
AI-Powered MCP Weather System
Open-sourced full-stack weather platform built on the Model Context Protocol with Open-Meteo MCP server and Streamlit client. Supports natural-language queries.
Mini-UAV Detection Publication
Paper published in Operations Research Forum. Achieved 90% precision on aerial imagery through advanced augmentation and hyperparameter tuning.
Foul Detection System
Real-time foul detection from UAV aerial streams in open zones using Dfine object detection combined with a clustering mechanism to identify and classify foul events with high precision.
Contact
Open to AI, CV, and MLOps roles
I’m available for projects in sports analytics, intelligent document analysis, edge AI, and applied research.